Depth Control of a Biomimetic Manta Robot via Reinforcement Learning

This paper proposes a model-free biomimetic manta robot depth control method based on reinforcement learning. Different from the traditional control method, the reinforcement learning method does not need to establish a mathematical model of the control object, and autonomously learns the control law through data training. Based on the classical Q algorithm, the state space, the action space, and reward function of the depth control of the bionic manta robot are designed. The state-action function is trained offline using the experience replay mechanism and random sampling strategy. Finally, the trained function is transplanted to the biomimetic manta robot prototype to establish a controller. The effectiveness of the proposed control method is verified by experiments.